acceptodds
Under review as a conference paper at ICLR 2027

Seek Before You Move: Evidence Seeking for Progress Grounding in Vision-Language Navigation

Abstract

Vision-Language Navigation (VLN) requires agents to continuously ground task progress from long-horizon instructions and partial egocentric observations. Existing VLM-based navigation agents typically reason only over available observations and may remain confident even when task-relevant evidence is missing. For example, an agent may confidently proceed forward and get lost even though the landmark indicating the next turn lies outside its current field of view. We term this failure mode Progress Myopia: the agent fails to recognize unreliable progress grounding and continues acting on insufficient evidence. To address it, we propose SeekVLN, an evidence-seeking framework that couples semantic progress reasoning with active acquisition of task-relevant observations. SeekVLN is trained in two stages: First, Future-guided Reverse Generation (FRG) uses future expert actions to augment offline expert trajectories with supplementary views and evidence annotations. Supervised fine-tuning on these trajectories establishes a prior for evidence seeking and progress reasoning without additional expert interaction. However, imitation alone does not reveal whether seeking improves subsequent navigation. We therefore introduce Counterfactual Contrastive Policy Optimization (C2PO) for reinforcement fine-tuning. By comparing each evidence-seeking branch with a counterfactual direct-navigation branch from the same state, C2PO uses a contrastive reward to assign credit to seeking decisions based on subsequent navigation benefit. Experiments on simulated benchmarks show that SeekVLN achieves state-of-the-art performance, improving success rate by 12.7% and 7.5% over the base model on R2R-CE and RxR-CE, respectively. Both simulated and real-world evaluations exhibit human-like evidence-seeking behaviors for more reliable progress grounding.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language Navigation, Progress Grouding, Reinforcement Learning, Vision-Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_D1hGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a framework called SeekVLN, which improves progress anchoring in visual-language navigation through active evidence seeking. The authors identify a failure mode known as “progress myopia,” in which agents make confident navigation decisions even when visual evidence is insufficient. SeekVLN employs a dual-mode strategy that allows agents to choose between direct navigation and acquiring supplementary observations before reasoning and action. The proposed training workflow includes Future-Guided Reverse Generation (FRG) that uses offline expert trajectories to construct supervision for evidence search and progress reasoning, and Counterfactual Comparison Strategy Optimization (C2PO) that estimates the benefits of evidence acquisition by comparing evidence-search branches with direct-navigation branches. Experiments on R2R-CE and RxR-CE demonstrate that this method achieves improvements over baseline models and previous competitive approaches. Additional ablation experiments and qualitative real-world demonstrations further validate the effectiveness of the proposed method.

Critical strengths
  1. The paper highlights an important limitation of navigation agents operating under partial observations. Allowing an agent to seek additional visual evidence before making uncertain decisions is a practically relevant direction for embodied navigation.

  2. C2PO provides a reasonable mechanism for assigning credit to evidence-seeking decisions based on subsequent navigation progress. The reported improvements are substantial, and the ablations indicate that the counterfactual reward contributes additional benefits beyond standard reinforcement fine-tuning.

Critical weaknesses
  1. Unclear whether the model genuinely learns to recognize insufficient visual evidence. As shown in Table 2 of Appendix B, FRG generates SEEK supervision using predefined proportions conditioned on expert actions. For example, 30% of forward actions and 75% of turns with a cumulative rotation of 45° are labeled as SEEK. Therefore, the initial SEEK supervision is constructed using action-dependent heuristics rather than directly assessing the sufficiency of available visual evidence. Moreover, the SEEK mode always acquires three fixed views (left, front, and right), without actively selecting where to observe. While the results demonstrate that the model can learn when to acquire supplementary views, they do not sufficiently establish that it genuinely recognizes insufficient visual evidence, which is a central claim of the paper.

  2. Limited flexibility and generality of the proposed evidence-seeking mechanism. Despite its reasonable motivation, the proposed framework relies on a relatively constrained evidence-seeking mechanism. Specifically, the model switches between direct navigation and a predefined three-view observation procedure, rather than adaptively determining which additional observations would be most informative. Furthermore, recent VLN approaches have increasingly emphasized more general instruction understanding and task-level reasoning, which may not require explicit landmark identification at every decision step. The paper does not sufficiently discuss how its evidence-seeking mechanism generalizes to these navigation paradigms. These limitations weaken the generality and methodological novelty of the proposed framework.

Questions for authors
  1. Generalizability across different VLN backbones. Is the proposed evidence-seeking framework applicable to other VLN architectures beyond Aux-Think? Why did the authors choose Aux-Think as the base model instead of stronger methods such as Progress-Think or StreamVLN, which achieve better performance in Table 1? Evaluating the framework on additional backbones would provide stronger evidence of its generalizability and contribution.

  2. Performance gains on more challenging navigation benchmarks. RxR-CE generally involves longer and more challenging navigation trajectories than R2R-CE, where reliable progress grounding is particularly important. However, the proposed method improves SR from 54.8% to 67.5% on R2R-CE, compared with 52.2% to 59.7% on RxR-CE. Could the authors explain the relatively smaller improvement on RxR-CE? More importantly, how much of the improvement comes from standard reinforcement fine-tuning, and how much can be specifically attributed to the proposed counterfactual evidence-seeking optimization? Although Table 6 provides a relevant ablation on R2R-CE, an equivalent analysis on RxR-CE would help clarify this distinction. This requires further research by the authors to determine whether the performance improvement can genuinely be attributed to the proposed evidence-seeking framework.

  3. Insufficient comparisons with recent VLN methods. Table 1 includes only one method from 2026, despite the emergence of several recent VLN approaches. In particular, the paper discusses methods such as AwareVLN and Dual-Anchoring in the related work but does not include them in the experimental comparisons. Could the authors provide comparisons with more recent and relevant methods under compatible evaluation settings, or explain why these methods were excluded? Such comparisons would help establish the competitiveness of SeekVLN against current VLN approaches.

I would currently lean toward a borderline reject, but I would be open to increasing the rating if the rebuttal provides clarifications on the questions and concerns.

7 comments
AuthorsOpenReview ↗

Consolidated Response to Reviewer D1hG (W1&W2)

Weakness 1: Insufficient evidence that SeekVLN has learned to recognize insufficient visual evidence.

“Unclear whether the model genuinely learns to recognize insufficient visual evidence. ... the initial SEEK supervision is constructed using action-dependent heuristics rather than directly assessing the sufficiency of available visual evidence. ... While the results demonstrate that the model can learn when to acquire supplementary views, they do not sufficiently establish that it genuinely recognizes insufficient visual evidence, which is a central claim of the paper.”

Response:

We agree that the original end-to-end results and trigger analyses alone do not establish recognition of insufficient evidence. We have therefore revised our diagnostic experiments, introducing Task-Relevant Evidence Diagnosis (TRED) and evidence-conditioned SEEK measurements.

In TRED, an expert VLM reasons over the instruction and observation history to label evidence-sufficient and evidence-insufficient states. We evaluate representative agents on these annotated expert trajectories. All three agents exhibit markedly higher action error rates when evidence is insufficient. Table R.1 shows that SeekVLN reduces errors relative to Aux-Think by 15.98 percentage points under insufficient evidence and 7.14 under sufficient evidence. These results appear in Figs. 1(c) and 3(a) of the revised manuscript, with the diagnostic design in Appendix A.

Table R.1. Action error rates (%) by evidence sufficiency.

AgentSufficient ↓Insufficient ↓Gap (pp)
NaVILA24.8637.9713.11
StreamVLN21.7630.428.66
Aux-Think34.9646.4211.46
SeekVLN27.8230.442.62

On the same offline expert-trajectory states as Table R.1, SeekVLN-SFT’s mean SEEK probability and selection rate increase by 10.79 and 17.07 percentage points under insufficient evidence (Table R.3). This supports its ability to proactively seek at evidence-insufficient states. This analysis is also included in Section 4.4 and Fig. 3(a) of the revised manuscript.

Table R.3. SEEK decisions of SeekVLN (%).

EvidenceMean SEEK prob.SEEK selection
Sufficient34.329.74
Insufficient45.1126.81
Gap (pp)+10.79+17.07

Gap = insufficient − sufficient; pp = percentage points. R.3 reports offline SFT decisions, with probabilities normalized over SEEK/NAV.

FRG’s action-conditioned sampling (Table 2, Appendix B of the original submission) primarily supplies a behavioral prior. C2PO’s counterfactual reward is the key optimization signal designed to train adaptive seeking, measuring its navigation benefit over a matched direct-navigation branch. Together, the new diagnosis and behavioral analyses support SeekVLN’s ability to identify evidence-insufficient states and proactively seek additional observations.

Weakness 2: Fixed three-view seeking limits flexibility and generality.

“Limited flexibility and generality of the proposed evidence-seeking mechanism. ... the model switches between direct navigation and a predefined three-view observation procedure, rather than adaptively determining which additional observations would be most informative. ... The paper does not sufficiently discuss how its evidence-seeking mechanism generalizes to these navigation paradigms.”

Response:

The three-view procedure is a specific implementation of evidence seeking; C2PO is not restricted to this design. With sufficient resources, evidence-seeking data could cover different viewing angles and numbers of views, active viewpoint selection, or exploratory movements. Given time and cost constraints, we supplement the current observation with left and right views to demonstrate evidence seeking for reliable progress grounding.

Regarding methods emphasizing general instruction understanding and task-level reasoning, these capabilities alone do not resolve Progress Myopia. Aux-Think explicitly uses task-progress reasoning, yet still exhibits markedly higher action errors under insufficient evidence (Table R.1). This difficulty stems from the physical limitation of partial observability, rather than simply limited reasoning or representation: stronger reasoning over existing observations cannot supply task-relevant visual evidence that has not been observed.

More broadly, C2PO could help VLN models learn proactive reasoning and exploration behaviors beyond evidence seeking. SeekVLN thus addresses an information limitation that understanding and reasoning alone cannot resolve, while C2PO has the potential to support more general proactive behavior learning.

AuthorsOpenReview ↗

Consolidated Response to Reviewer D1hG (Q2)

Question 2: Smaller RxR gains and the contribution beyond standard RL.

Response:

Thank you for raising these questions. SeekVLN’s RxR-CE results are obtained without any RxR training, and our ablations show that counterfactual optimization improves over standard PPO on both benchmarks. We have clarified this transfer setting and added the corresponding RxR-CE reward ablations.

Q2.1 RxR-CE is evaluated without RxR training.

"Performance gains on more challenging navigation benchmarks. RxR-CE generally involves longer and more challenging navigation trajectories than R2R-CE, where reliable progress grounding is particularly important. However, the proposed method improves SR from 54.8% to 67.5% on R2R-CE, compared with 52.2% to 59.7% on RxR-CE. Could the authors explain the relatively smaller improvement on RxR-CE? "

We believe that training exclusively on R2R-CE is the main reason for the smaller gains on RxR-CE. Neither SFT nor RFT uses any RxR data, so RxR-CE evaluation is an out-of-distribution, zero-shot transfer test. Nevertheless, with only approximately 4.8K R2R-CE training episodes across the two stages, SeekVLN improves RxR-CE SR by 7.5 percentage points over Aux-Think. This gain supports the effectiveness of evidence seeking for progress reasoning and grounding in longer-horizon navigation tasks.

Q2.2 C2PO improves over standard PPO.

"More importantly, how much of the improvement comes from standard reinforcement fine-tuning, and how much can be specifically attributed to the proposed counterfactual evidence-seeking optimization? "

The w/o CF Reward variant in Table 6 of the original submission is standard PPO fine-tuning: it retains the FRG-SFT initialization and outcome reward but removes the counterfactual reward. This comparison therefore isolates C2PO’s additional contribution beyond standard reinforcement fine-tuning. As shown in Table R.2, Full C2PO improves SR and SPL over this variant by 3.43 and 1.96 percentage points, respectively. The w progress reward control additionally rewards the SEEK branch’s own progress without counterfactual comparison; Full C2PO improves SR/SPL by 2.13/2.66 points over this control. These results are reported in Section 4.3, Table 3 of the revised manuscript.

Table R.2. C2PO reward ablations on R2R-CE Val-Unseen-613 (%).

VariantSR ↑SPL ↑
SeekVLN-FRG-SFT60.7055.80
w/o CF Reward65.0960.40
w progress reward66.3959.70
Full C2PO68.5262.36

Q2.3 The benefit also transfers to RxR-CE.

"Although Table 6 provides a relevant ablation on R2R-CE, an equivalent analysis on RxR-CE would help clarify this distinction. This requires further research by the authors to determine whether the performance improvement can genuinely be attributed to the proposed evidence-seeking framework."

The added RxR-CE ablations show that C2PO’s benefit over standard PPO generalizes beyond the training benchmark. All variants are trained on R2R-CE and evaluated on RxR-CE Val-Unseen without RxR training. As shown in Table R.4, Full C2PO improves SR and SPL over w/o CF Reward by 2.59 and 0.64 percentage points, respectively, and achieves the best SR, SPL, and NE among the tested variants.

Table R.4. Reward ablations under zero-shot transfer to RxR-CE.

VariantSR (%) ↑SPL (%) ↑NE ↓
SeekVLN-SFT55.7047.405.8
w/o CF Reward57.1049.705.0
w progress reward57.7550.155.1
Full C2PO59.6950.344.9

Thank you for the detailed rebuttal and substantial additional experiments. I appreciate the efforts to address my concerns. In particular, the budget-matched triggering comparisons and the new RxR-CE reward ablations provide stronger evidence for the effectiveness of adaptive seeking and the additional contribution of C2PO over standard PPO.

However, several key concerns remain before increasing the rating:

Does the proposed SeekVLN genuinely recognize insufficient visual evidence? The new TRED analysis is valuable, and the higher SEEK probability under evidence-insufficient states is encouraging. Nevertheless, the original FRG supervision remains strongly dependent on expert action types, while evidence insufficiency may also correlate with turning actions and difficult decision points. Therefore, the observed correlation does not fully establish that the model learns evidence sufficiency rather than action-dependent seeking priors. Could the authors report SEEK probabilities or selection rates under sufficient and insufficient evidence within the same expert-action categories? Additionally, Table R.3 evaluates the SFT model rather than the final C2PO-trained policy. An equivalent analysis of the final policy would directly strengthen the central claim.

Reviewer_bdbUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes SeekVLN, an evidence-seeking framework for Vision-Language Navigation (VLN). The primary focus is dealing with "progress myopia", wherein agents make overconfident decisions when their progress grounding is unreliable due to insufficient observation. The authors introduce a two-stage learning pipeline: (1) Future-guided Reverse Generation (FRG) to construct evidence-seeking trajectories, and (2) Counterfactual Contrastive Policy Optimization (C2PO) to assign credit via reinforcement learning. Experiments on R2R-CE and RxR-CE benchmarks show state-of-the-art numbers compared to recent Vision-Language Model (VLM) approaches, and analysis claims SeekVLN acquires human-like evidence-seeking behaviors.

Critical strengths
  1. The authors attempt to formalize and attack a less-discussed VLN failure mode (“progress myopia”), raising awareness of confidence calibration and evidence acquisition in embodied navigation.

  2. The core training loop (FRG->C2PO) is a novel application of counterfactual reinforcement learning in VLN, and could inspire future work.

Critical weaknesses
  1. All ablation experiments (including key questions about the individual effects of FRG and C2PO) are moved out of the main text. Because these are critical to understanding whether each module of SeekVLN is necessary or effective, this is not a reasonable format for a conference paper—the arguments about the validity and necessity of the method are not directly accessible to reviewers and readers. Fixing this would require a major restructuring of the manuscript.

  2. The authors do not include any experimental comparison with the recent method MapDream (ICML 2026) . This makes it hard to judge whether the proposed approach is genuinely state-of-the-art or simply orthogonal to existing baselines.

  3. The presentation is not self-contained, the Discussion and Results sections are overly brief regarding negative cases, failure analysis, and behavioral differences with related work.

  4. Since the base model is auxthink, it would be better if the authors could show how each design improves auxthink,

Questions for authors

see weakness

4 comments
AuthorsOpenReview ↗

Consolidated Response to Reviewer bdbU

Concern 1: The ablation studies should be accessible in the main text.

All ablation experiments (including key questions about the individual effects of FRG and C2PO) are moved out of the main text. Because these are critical to understanding whether each module of SeekVLN is necessary or effective, this is not a reasonable format for a conference paper. ...

Response:

We have moved the FRG and C2PO ablations into Section 4.3, “Ablation Studies,” between “Main Results” and “Deep Dive into Evidence-Seeking Behavior.” This makes the contribution of each component directly accessible in the main text.

Our original arrangement prioritized evidence-seeking analysis because it directly examines our central contribution—whether the agent learns when to seek and benefits from the acquired evidence.

Concern 2: The main comparison omits MapDream.

The authors do not include any experimental comparison with the recent method MapDream (ICML 2026) . This makes it hard to judge whether the proposed approach is genuinely state-of-the-art or simply orthogonal to existing baselines.

Response:

Thank you for this suggestion. We have added MapDream to the main table.

Our study focuses on progress grounding in VLN, so we include two groups of baselines: established general-purpose VLN agents, including NaVid, Uni-NaVid, NaVILA, NavFoM, StreamVLN, and now MapDream; and methods emphasizing progress reasoning and task grounding, including Progress-Think, Aux-Think, and now Dual-Anchoring. MapDream jointly learns navigation-oriented bird’s-eye-view map generation and action prediction, using compact maps to retain task-relevant spatial context. We therefore include it in the first group.

Concern 3: Negative cases, failure analysis, and behavioral comparisons need more discussion.

The presentation is not self-contained, the Discussion and Results sections are overly brief regarding negative cases, failure analysis, and behavioral differences with related work.

Response:

We have added a qualitative failure-case comparison with Progress-Think and Aux-Think in the appendix (“Qualitative Comparison with Baselines”). In this example, both baselines miss the instruction to go upstairs and remain on the entry level, whereas SeekVLN’s additional left view reveals the staircase and enables it to follow the upstairs route, illustrating how evidence seeking supports progress grounding at a critical route choice.

Concern 4: The contribution of each design over Aux-Think needs to be shown.

Since the base model is auxthink, it would be better if the authors could show how each design improves auxthink.

Response:

Our FRG ablations isolate how evidence-seeking supervision and progress reasoning each improve upon Aux-Think during SFT. These results, originally in Table 5, now appear in Table 2 of the revised main text.

Table 2. FRG ablations on R2R-CE Val-Unseen-613 (NE in meters; SR/SPL in %).

VariantNE ↓SR ↑SPL ↑
Aux-Think (base model)5.8052.045.0
w/o Evidence Seeking4.8658.653.3
w/o Progress Reasoning4.9958.753.0
Full FRG4.8460.755.8

On R2R-CE Val-Unseen-613, Full FRG improves SR from 52.0% to 60.7% and SPL from 45.0% to 55.8% over Aux-Think. Removing evidence-seeking supervision reduces SR/SPL to 58.6%/53.3%, while removing progress reasoning reduces them to 58.7%/53.0%. Both variants underperform Full FRG, showing that each supervision component contributes to the improvements over the base model.

Thanks. The new FGR results is inconsistency with the results in Table 1. Why?

AuthorsOpenReview ↗

Follow-up: Different evaluation sets for Table 1 and the FRG ablations.

Thank you for pointing this out! The difference arises because Table 1 and the FRG ablations use different evaluation sets. Table 1 reports results on the full R2R-CE Val-Unseen split, whereas the ablations use a trajectory-deduplicated subset of 613 episodes, retaining one instruction per unique expert trajectory.

We use this subset to reduce evaluation costs, as a single evaluation on the full split typically takes 12-16 hours. Table R.5 below shows that both evaluation sets yield the same model ranking across SR, NE, and SPL, with consistent improvements from Aux-Think to SeekVLN-SFT and then SeekVLN-RFT. This supports the subset’s use for controlled comparisons, although the absolute scores differ.

Table R.5 Comparison between the full R2R-CE Val-Unseen split and the 613-episode subset.

Val-UnseenVal-Unseen-613
ModelNE ↓SR ↑SPL ↑NE ↓SR ↑SPL ↑
Aux-Think6.0854.8046.905.8052.0045.00
SeekVLN-SFT4.7061.0055.904.8460.7055.80
SeekVLN-RFT3.7067.5061.403.6668.5262.36

All ablation variants are evaluated on the same subset to ensure a fair comparison. We have clarified the rationale and details of this evaluation setting in Lines 383–385 of the revised manuscript.

AuthorsOpenReview ↗

Follow-up response to reviewer bdbU: Full-split ablations

Follow-up concern: Thanks. I need the full results before rasing my score. And the code should be released upon acceptance for verifing this since the initial inconsistency.

Thank you for your follow-up! We have now completed both the FRG and C2PO ablations on the full R2R-CE Val-Unseen split, reported in Tables R.6 and R.7 below. We have also replaced the 613-episode subset results in Section 4.3, Tables 2 and 3 of the revised manuscript with these full-split results, so the ablations and main results now use the same evaluation split.

The full-split results support the same conclusions as the subset experiments. For FRG, removing either evidence-seeking supervision or progress reasoning lowers SR and SPL and increases NE; combining both achieves the best performance. Relative to Full FRG, removing evidence seeking reduces SR/SPL by 3.92/4.57 percentage points, while removing progress reasoning reduces them by 4.35/5.19 points.

Table R.6. FRG ablations on the full R2R-CE Val-Unseen split.

VariantNE (m) ↓SR (%) ↑SPL (%) ↑
Aux-Think (base model)6.0854.8046.90
w/o Evidence Seeking4.9657.0451.35
w/o Progress Reasoning5.1256.6150.73
Full FRG4.7260.9655.92

For C2PO, removing the counterfactual reward reduces SR/SPL by 2.57/1.43 points. Retaining the dense progress reward partially recovers SR, but slightly lowers SPL relative to w/o CF Reward. Full C2PO still improves SR/SPL over this control by 1.05/1.57 points and achieves the lowest NE. Thus, counterfactual comparison contributes beyond a dense progress reward, supporting its effectiveness in learning beneficial adaptive evidence seeking.

Table R.7. C2PO ablations on the full R2R-CE Val-Unseen split.

VariantNE (m) ↓SR (%) ↑SPL (%) ↑
SeekVLN-FRG-SFT4.7260.9655.92
w/o CF Reward3.9364.9359.97
w progress reward3.7566.4559.83
Full C2PO3.7067.5061.40

We will release the code and corresponding checkpoints upon acceptance, including the evaluation and ablation configurations, to enable verification.

Reviewer_gF7sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of unreliable progress grounding in Vision-Language Navigation (VLN). The authors identify a failure mode termed Progress Myopia, in which navigation agents remain confident in their decisions even when the available visual evidence is insufficient for reliable progress estimation. By comparing evidence-seeking and direct-navigation branches initialized from the same state, the method provides a reward signal based on their short-horizon navigation progress differences. The proposed method is evaluated on R2R-CE and RxR-CE.

Critical strengths
  1. The two-stage training framework is reasonably motivated and technically coherent. The combination of FRG and C2PO provides a reasonable approach to learning evidence-seeking behavior.

  2. The method demonstrates substantial improvements on established VLN benchmarks.

Critical weaknesses
  1. The formulation of Progress Myopia is central to the motivation of the proposed method. However, the evidence presented in Figure 1(c) is insufficient to establish that existing navigation agents systematically fail to recognize unreliable progress grounding. Similar confidence values between successful and failed segments may indicate poor calibration or limited discriminative power of the chosen confidence metric. They do not directly establish that the agent is unable to recognize missing task-relevant visual evidence.

  2. The ablations do not fully isolate the contribution of counterfactual credit assignment. The counterfactual contrastive reward is a major methodological contribution of the paper. The authors compare SEEK and NAV branches from the same state and use their difference in short-horizon geodesic progress as an additional training reward. However, the main C2PO ablation removes the counterfactual reward entirely while retaining the terminal outcome reward. Therefore, this comparison simultaneously changes two aspects of the training signal: whether the reward is counterfactual and whether the policy receives an additional dense progress-related reward.

  3. The proposed evidence-seeking mechanism is limited to a fixed observation pattern, and its acquisition cost is insufficiently considered.

Questions for authors

See weaknesses.

  1. The typo “forawrd” should be corrected to “forward”
3 comments
AuthorsOpenReview ↗

A kind follow-up on our response

Dear Reviewer gF7s,

We sincerely thank the reviewer again for the time and care devoted to our paper. As the discussion period is coming to an end, we would like to kindly ask whether our response has addressed the reviewer's concerns. We have revised the paper accordingly, with all changes marked in blue. If any points remain unclear, we would be glad to provide further clarification before the discussion closes.

We are very grateful for the reviewer's thoughtful feedback.

Best regards, The Authors

AuthorsOpenReview ↗

Updated Consolidated Response to Reviewer gF7s ( W1)

We have conducted more comprehensive experiments to further examine the motivation for SeekVLN and have updated our response to Concern 1 accordingly.

Concern 1: The confidence-based diagnosis does not establish failures to recognize insufficient evidence.

Weakness 1: ... However, the evidence presented in Figure 1(c) is insufficient to establish that existing navigation agents systematically fail to recognize unreliable progress grounding. ... They do not directly establish that the agent is unable to recognize missing task-relevant visual evidence.

Response:

We agree that confidence alone does not establish a link between insufficient evidence and unreliable progress grounding. We have refined the definition of Progress Myopia and replaced the confidence-based analysis with Task-Relevant Evidence Diagnosis (TRED), updating the Abstract, Introduction, Figure 1, and Appendix A. TRED uses an expert VLM to label evidence sufficiency from the instruction and observation history. We then compare representative agents’ action error rates across the two groups of annotated expert-trajectory states (Table R.1).

Table R.1. Action error rates (%) by evidence sufficiency.

AgentSufficient ↓Insufficient ↓Gap (pp)
NaVILA24.8637.9713.11
StreamVLN21.7630.428.66
Aux-Think34.9646.4211.46
SeekVLN-SFT27.8230.442.62
SeekVLN-RFT22.9524.061.11

Gap = insufficient − sufficient; pp = percentage points.

All three representative agents make markedly more action errors under insufficient evidence, supporting our observation of Progress Myopia: agents struggle to maintain reliable progress grounding when evidence is insufficient.

Both SeekVLN-SFT and SeekVLN-RFT reduce action errors relative to Aux-Think in both evidence groups. SeekVLN-RFT reduces errors by 12.01 percentage points under sufficient evidence and 22.36 points under insufficient evidence, supporting evidence seeking as a way to mitigate Progress Myopia.

On the same offline expert-trajectory states as Table R.1, both SeekVLN-SFT and SeekVLN-RFT show higher mean SEEK probabilities and selection rates under insufficient evidence (Table R.3). After C2PO fine-tuning, the insufficient-minus-sufficient gaps increase from 10.79 to 16.19 percentage points for mean SEEK probability and from 17.07 to 19.89 points for selection rate, supporting stronger evidence-conditioned seeking in the final policy.

Table R.3. SEEK decisions of SeekVLN-SFT and SeekVLN-RFT (%).

EvidenceSFT Mean SEEK prob.SFT Selection rateRFT Mean SEEK prob.RFT Selection rate
Sufficient34.329.7427.0311.84
Insufficient45.1126.8143.2231.74
Gap (pp)+10.79+17.07+16.19+19.89

Gap = insufficient − sufficient; probabilities are normalized over SEEK/NAV. Selection rate is the fraction choosing SEEK under this binary decision.

The revised definition is marked in blue in the Abstract and Introduction, alongside a clearer motivation. We also added the Mitigating Progress Myopia analysis in Section 4.4 (Fig. 3(a)), including the SEEK behavior results in Table R.3.

AuthorsOpenReview ↗

Consolidated Response to Reviewer gF7s (W2&W3)

Concern 2: The ablations do not separate counterfactual comparison from dense progress rewards.

Weakness 2: The ablations do not fully isolate the contribution of counterfactual credit assignment. ... Therefore, this comparison simultaneously changes two aspects of the training signal: whether the reward is counterfactual and whether the policy receives an additional dense progress-related reward.

Response:

We agree that an additional control is needed and have added w progress reward to isolate counterfactual credit assignment. It retains the outcome reward and the SEEK branch’s own short-horizon progress reward without counterfactual comparison. Full C2PO improves SR/SPL over this control by 2.13/2.66 percentage points (Table R.2). Adding progress reward alone improves SR but lowers SPL relative to w/o CF Reward. Thus, counterfactual comparison provides gains beyond an additional dense progress reward.

Table R.2. C2PO reward ablations on R2R-CE Val-Unseen-613 (%).

VariantSR ↑SPL ↑
SeekVLN-FRG-SFT60.7055.80
w/o CF Reward65.0960.40
w progress reward66.3959.70
Full C2PO68.5262.36

Concern 3: The fixed observation pattern and acquisition cost need further justification.

Weakness 3: The proposed evidence-seeking mechanism is limited to a fixed observation pattern, and its acquisition cost is insufficiently considered.

Response:

We agree that the observation pattern and acquisition cost need clarification. Fixed left/right views are a practical implementation choice for demonstrating evidence seeking, rather than an inherent restriction of C2PO. The framework could accommodate other viewing angles, numbers of views, or exploratory movements.

Side views can be obtained through additional side-facing cameras in simulation or lateral cameras on robots, motivating our lightweight fixed-view implementation. Physical acquisition cost depends on the robot’s sensing and motion configuration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.